This is a complete, unsliced EXL3 K2 checkpoint of
deepseek-ai/DeepSeek-V4-Pro-0813.
It carries the original dense/shared weights and K2 trellis encodings for every
routed expert in all 61 base layers and all three native dSpark blocks. The
standard Hugging Face artifact contains 297,238 tensors in 51 shards, with
407.57 GiB of tensor payload. config.json carries only the compact EXL3
format declaration; the full per-tensor storage map is kept separately in
quantize_config.json.
Serve this checkpoint with DS4RT
The complete inference engine, five-host launcher, native source, architecture
documentation, reproducible build, and public benchmark record are available
in tpurtell/ds4rt-pro-rtx-4spark.
The supported release topology is one NVIDIA RTX PRO 6000 Blackwell 96 GB
coordinator plus four NVIDIA DGX Spark expert workers on a dedicated 400-Gb/s
path.
For the published v1 coordinator and Spark images, start with the
repository's quick-deploy guide:
bash
1git clone --recurse-submodules \2 https://github.com/tpurtell/ds4rt-pro-rtx-4spark.git
3cd ds4rt-pro-rtx-4spark
45# Populate this exact model revision on all five hosts, then configure the6# four Spark host/address pairs in ds4rt.config.7./run.sh --dry-run
8./run.sh
The encoder replayed a frozen, deterministic 1,426-prompt corpus containing
about 1.081 million English-, Chinese-, code-, reasoning-, instruction-, and
structured-output-focused tokens through the real learned routers. Naturally
under-covered experts were augmented deterministically from router ranks 7–12
to a 1,024-row target, with the declared isotropic recovery residual used only
when that natural tail was exhausted. Each dSpark block used a deterministic
327,680-anchor sample of the same activation stream.
The run was streamed layer by layer with bounded two-RTX Hessians, resumable
batch capture/replay, per-projection numerical validation, and a retained
per-layer/expert error and coverage ledger. No Spark compute was used during
quantization.
Validation
The final checkpoint passed its complete tensor/shape/digest contract and was
served on the open-source
DS4RT Pro v1 engine: one
RTX PRO 6000 Blackwell coordinator on a 400-Gb/s link plus four DGX Spark TP4
expert ranks, using balanced FP8 KV and adaptive dSpark.
Post-fix gate
Result
Tool Eval Bench 2.3.2, all 69 scenarios, no thinking
Repeated-prefix retained 32K + 32K prefill, median over two
2,039.0 tok/s
Adaptive dSpark acceptance on the semantic blend
3,568/5,577 (63.98%)
These are August 21 post-attention-fix measurements. Earlier Tool Eval figures
on this card were invalidated by the attention defect and have been replaced,
not combined with the corrected run. The 68.09 tok/s result asks for orchid
exactly 100 times; the 2,039.0 tok/s retained result extends a repeated-alpha
prefix and excludes prefix construction. They are explicit low-entropy
ceilings rather than semantic-workload claims.
Tool Eval's safety gate did not pass: it flagged TC-34, TC-42, TC-58, and
TC-60. Treat the score as tool-use quality evidence, not a safety
qualification. All 69 scenarios completed without an HTTP/backend failure.
The complete nine-cell C1/C2/C4 decode matrix (27 timed runs) and all 63
requests finished without a serving failure. Every timed repeat-decode and
prefill sample completed full attention and numeric progression with zero
request-time CUDA graph captures.
2K prompt + 128 generated tokens
C1
C2 aggregate
C4 aggregate
0 context
37.6
49.0
54.2
4K context
30.2
44.6
53.8
8K context
29.7
47.5
50.6
The full fresh/retained prefill matrix, per-cell measurement contract, and
provenance are kept in the linked public benchmark record rather than copied
into this weight repository.
The original model card follows for architecture, prompting, benchmark, and
license information. Its native-weight vLLM/SGLang launch examples are not
EXL3 serving instructions for this checkpoint.
DeepSeek-V4-Pro-0813 is the official release of DeepSeek-V4-Pro, superseding the preview version, with greatly enhanced agentic capabilities and performance improvements that are especially pronounced in production environments. It is built on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative decoding module attached.
DeepSeek-V4-Pro-0813 outperforms DeepSeek-V4-Pro (Preview) on the benchmarks listed below, and is broadly competitive with the strongest proprietary models available.
Benchmark
DeepSeek-V4-Pro-0813
DeepSeek-V4-Flash-0731
DeepSeek-V4-Pro (Preview)
DeepSeek-V4-Flash (Preview)
GLM-5.2
Kimi K3
Opus-4.8
Fable-5 (w/ fallback)
HLE (wo / w tools)
42.7 / 60.0
37.8 / 51.5
37.7 / 48.2
34.8 / 45.1
40.5 / 54.7
43.5 / 56.0
49.8 / 57.9
53.3 / 63.0
Terminal Bench 2.1
87.9
82.7
72.1
61.8
81.0
88.3
85.0
88.0
NL2Repo
61.5
54.2
38.5
39.4
48.9
-
69.7
-
Cybergym
83.3
76.7
52.7
38.7
-
80.0
78.3
83.1
DeepSWE
62.7
54.4
12.8
7.3
46.2
67.5
58.0
70.0
Toolathlon-Verified
74.1
70.3
55.9
49.7
59.9
76.5
76.2
77.9
Agents' Last Exam
25.7
25.2
16.5
15.8
23.8
27.6
25.7
-
AutomationBench (Public)
31.8
25.1
12.8
10.8
12.9
30.8
27.2
29.1
DSBench-FullStack †
71.1
68.7
41.8
37.0
61.8
73.7
71.6
77.2
DSBench-Hard †
67.2
59.6
31.1
25.8
54.5
63.0
71.7
68.3
Notes:
For the code-agent tasks among the public benchmarks above, DeepSeek-V4-Pro-0813 is evaluated with the minimal mode of DeepSeek Harness as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95.
† DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.
Chat Template
This release does not include a Jinja-format chat template. Instead, we provide a dedicated encoding folder with Python scripts and test cases demonstrating how to encode messages in OpenAI-compatible format into input strings for the model, and how to parse the model's text output. Please refer to the encoding folder for full documentation.
The reasoning_effort parameter now supports three levels — low, high, and max — which control how much deliberation the model spends before answering.
For example, the command below serves the model with vLLM on a single 4×GB300 node.
See the vLLM recipe for detailed instructions and other hardware configurations.
Enable DSpark with --speculative-algorithm DSPARK and do not set a separate --speculative-draft-model-path as the target and draft weights therefore come from the same checkpoint.
See the SGLang cookbook for detailed instructions, benchmarks and other hardwares configurations.
Please refer to the inference folder for detailed instructions on running DeepSeek-V4 locally, including model weight conversion and interactive chat demos.
For local deployment, we recommend setting the sampling parameters to temperature = 1.0, with top_p = 0.95 for agentic scenarios and top_p = 1.0 otherwise. For the high and max reasoning effort levels, we recommend a maximum output length of 384K tokens.
License
This repository and the model weights are licensed under the MIT License.
